AI Agents
Create AI-powered voice agents that handle natural conversations with callers, understand context, and respond intelligently within your IVR flows.
Overview
AI Agents are configurable voice bots that use AI language models (OpenAI or Gemini) to process caller input in real-time. Speech-to-text converts the caller’s words to text, the LLM generates a response, and text-to-speech delivers it back to the caller. Agents can handle inquiries, provide information, qualify leads, and transfer calls to human agents when needed.
Creating an AI Agent
Step-by-Step
- Navigate to Conversational AI → AI Agents
- Click Create Agent
- Enter an Agent Name (descriptive, e.g., “Support Agent” or “Sales Qualifier”)
- Configure the Welcome Message — the first words callers hear
- Set the System Prompt — detailed instructions that define the agent’s behavior
- Choose the AI Model (OpenAI or Gemini models)
- Save the agent
Writing Effective System Prompts
The system prompt is the most critical configuration for your AI Agent. It defines the agent’s personality, scope of knowledge, tone, and behavioral boundaries. The model reads the system prompt before every interaction.
Prompt Engineering Best Practices. A well-written system prompt makes the difference between a helpful agent and a frustrating one. Follow these guidelines:
- Be specific about the role — e.g., “You are a helpful customer support agent for an e-commerce company named Acme Store.”
- Define the scope — clearly state what the agent can and cannot do
- Set the tone — professional, friendly, casual, or formal based on your brand
- Provide step-by-step instructions for common scenarios (greetings, FAQs, complaints)
- Include escalation rules — when and how to transfer to a human
- Keep it concise but comprehensive — long prompts increase latency and cost
System Prompt Example: Customer Support
You are a customer support agent for AcmeStore, an online electronics retailer.
Your role is to help customers with order inquiries, returns, and product questions.
Guidelines:
- Greet the caller warmly and ask how you can help.
- For order status: ask for the order number. Tell the customer you are looking it up.
- For returns: explain the 30-day return policy and generate a return request.
- For product questions: provide accurate information from your knowledge base.
- Do NOT make promises about refund amounts or delivery dates.
- If the caller is angry or demands to speak to a manager, apologize politely and transfer them to a human agent.
- Keep responses concise and under 20 seconds when spoken.
Tone: Professional, empathetic, and efficient.System Prompt Example: Lead Qualification
You are a sales qualification agent for Voxvaani, a business communication platform.
Your goal is to qualify inbound leads before routing them to a sales representative.
Guidelines:
- Introduce yourself and ask about the caller's business needs.
- Determine: company size, industry, budget range, and timeline.
- If the caller meets qualification criteria (50+ employees, immediate need), offer to connect them to a sales rep.
- If the caller is not qualified, provide basic information and direct them to the website.
- Do NOT discuss pricing — let the sales rep handle that.
- Collect the caller's name, company, and best callback number.
Tone: Friendly, consultative, not pushy.Welcome Message
The welcome message is the first thing callers hear when they reach your AI Agent. It sets expectations and guides the conversation.
Welcome Message Options
- Auto-generated — The system creates a welcome message based on your system prompt
- Custom written — Write your own message for complete control over the greeting
- Regenerated — Use the regenerate button to get a new auto-generated version
Welcome Message Examples
| Style | Example |
|---|---|
| Professional | ”Thank you for calling AcmeStore. I am an automated assistant here to help with your order inquiries and product questions. How can I help you today?” |
| Friendly | ”Hey there! You have reached AcmeStore support. Let me know what you need help with — orders, returns, or anything else — and I will get you sorted.” |
| Minimal | ”AcmeStore support. How can I help?” |
Keep Welcome Messages Short. Callers decide within the first 5-10 seconds whether to continue or hang up. Avoid lengthy introductions. State who you are, what you can help with, and prompt the caller to speak — all within 10-15 seconds.
AI Model Comparison
| Feature | OpenAI (GPT-4o / GPT-4o-mini) | Google Gemini |
|---|---|---|
| Latency | ~1-3 seconds per response | ~1-3 seconds per response |
| Cost | Higher per-token cost, especially GPT-4o | Generally lower cost per token |
| Voice Quality | Excellent natural language understanding | Strong multilingual support |
| Strengths | Nuanced conversation, strong instruction following | Cost-effective for high-volume use, strong at structured tasks |
| Best For | Complex customer interactions requiring subtlety | High-volume FAQ automation, simple routing |
Model Selection Strategy. For agents handling complex queries or requiring nuanced conversation (e.g., sales qualification, complaint handling), prefer OpenAI. For high-volume, straightforward use cases (e.g., store hours, order status), Gemini provides a more cost-effective option. Test both with your specific prompt to compare quality.
Testing Before Deployment
Before connecting an AI Agent to a live phone number, test it thoroughly:
- Use the built-in test console — simulate conversations and review responses
- Test common scenarios — greetings, FAQs, edge cases, angry customers
- Test fallback paths — what happens when the agent cannot understand the caller
- Test transfers — verify that escalations route to the correct human agent or IVR node
- Use a staging number — deploy to a test phone number and run real-world call tests with team members
Agent Fallback Behavior
When an AI Agent cannot determine the caller’s intent:
- The agent may ask clarifying questions (e.g., “I did not quite catch that — could you rephrase?”)
- After a configurable number of failed attempts, the agent triggers a fallback action
- Common fallback actions: transfer to a human agent, route to a standard IVR menu, or play a message with alternative contact methods
Always Configure Fallback Behavior. An agent without a fallback path can trap callers in an endless loop of confusion. Always define what happens when the agent reaches its understanding limit — a transfer to a real person is almost always the best option.
Using AI Agents in IVR Flows
AI Agents are added to IVR Flows as a node type. When a call reaches an AI Agent node, the system transfers control to the AI for natural conversation. Configure exit paths from the agent node for transfers, call termination, or routing to other IVR nodes.
Tips
- Write a clear, role-specific system prompt that defines boundaries and tone
- Test your agent with a variety of conversation scenarios before deploying
- Start with a narrow, specific use case and expand the agent’s scope over time
- Monitor conversation logs regularly to identify areas where the agent struggles
- Keep the welcome message under 15 seconds to reduce caller drop-off